Sign up

coolthor/gemma-4-12B-it-FP8-dynamic

🤗 On Hugging Faceany-to-anyapache-2.012B params13 GBsafetensors✓ HF checksums availableupdated today
Magnet

Gemma 4 12B-it — FP8 dynamic

![Buy Me a Coffee](https://buymeacoffee.com/thorlin)

Self-quantized FP8 (dynamic) of google/gemma-4-12B-it — Google's encoder-free omni model (text + image + audio + video). Quantized and benchmarked on an NVIDIA DGX Spark (GB10, sm_121a).

TL;DR: 13 GB on disk (from 23 GB BF16), 15.9 tok/s on a GB10 via vLLM, all four modalities intact. Data-free — no calibration needed.

Benchmark (GB10 / DGX Spark, vLLM 0.22.1 native, single-stream decode, warm)

| Format | Disk | tok/s (EN/ZH) | Omni |

|---|---|---|---|

| BF16 | 23 GB | 7.7 | yes |

| FP8 dynamic (this) | 13 GB | 15.9 | yes |

| NVFP4 W4A16 | 7.7 GB | 24.9 | yes |

If you want the smallest + fastest build, see the sibling NVFP4 weight-only repo. FP8 is the conservative choice (dynamic activations, no calibration, widest kernel support).

Accuracy (MMLU + TMMLU+) — near-lossless on both languages

I scored all three formats on MMLU (English, 57 subjects) and TMMLU+ (Traditional Chinese, 66 subjects) with lm-evaluation-harness, 5-shot, chat template applied, limit=30 (N ≈ 1,710 EN / 1,980 TC, ±~1.0 pt), through transformers:

| Format | MMLU (EN) | TMMLU+ (TC) | EN drop | TC drop |

|---|---|---|---|---|

| BF16 | 78.30% | 47.21% | — | — |

| FP8 dynamic (this) | 77.95% | 46.97% | −0.35 | −0.24 |

| NVFP4 W4A16 | 75.56% | 41.24% | −2.74 | −5.97 |

FP8 is the accuracy-preserving choice. Near-lossless on both languages (within ~0.4 pt) and symmetric — while weight-only NVFP4 drops Traditional Chinese by ~6 points. If you can spare the extra disk/bandwidth over NVFP4 and care about non-English quality, FP8 is the safer pick. limit=30, single model — indicative. Full writeup.

Quantization recipe

llmcompressor, scheme FP8_DYNAMIC, data-free (no calibration data). Ignore list keeps the head and the multimodal projectors in BF16:

QuantizationModifier(targets="Linear", scheme="FP8_DYNAMIC",
    ignore=["lm_head", "re:.*embed_vision.*", "re:.*embed_audio.*"])

Serving (vLLM)

Needs vLLM with native Gemma4UnifiedForConditionalGeneration (~0.22.x / main) and the TRITON_ATTN backend (Gemma 4 has heterogeneous head dims: head_dim 256 x 16 = 4096 != hidden 3840):

VLLM_ATTENTION_BACKEND=TRITON_ATTN \
vllm serve coolthor/gemma-4-12B-it-FP8-dynamic --max-model-len 4096

Environment (exact versions — this model is version-sensitive)

This is a brand-new arch on a brand-new GPU, so the toolchain matters more than usual. The versions I actually ran:

| Component | Version | Why it matters |

|---|---|---|

| vLLM | 0.22.1rc1.dev124 (main, post-PR) | Needs the native Gemma4UnifiedForConditionalGeneration class, which only landed around 0.22.x/main. On an older vLLM it falls back to the generic transformers backend, which mishandles Gemma 4's non-square attention and crashes on o_proj. |

| transformers | 5.10.1 | First release that knows model_type: gemma4_unified. Older transformers can't even load the config. |

| torch | 2.11.0+cu130 | The one that bit me. vLLM main pins torch==2.10, but its _C.abi3.so was compiled against 2.11+cu130 — installing the pinned 2.10 (and pip silently pulling the CPU wheel on arm64) gives an undefined symbol import error and a CPU-only build. Force-align: pip install --force-reinstall --no-deps torch==2.11.0 --index-url https://download.pytorch.org/whl/cu130. |

| compressed-tensors | bundled with llmcompressor | Reads the FP8 weight format. |

| GPU / arch | DGX Spark GB10, sm_121a, CUDA 13.x | The torch-ABI dance above is specific to building vLLM from source for sm_121. |

| Attention backend | VLLM_ATTENTION_BACKEND=TRITON_ATTN | Required, not optional — see the head-dim note above. |

On a normal CUDA GPU (Hopper/Ada/Blackwell desktop) you don't need the torch-ABI overlay — that pain is specific to building vLLM from source for sm_121. A recent pip install vllm (with the native Gemma4Unified class) plus VLLM_ATTENTION_BACKEND=TRITON_ATTN is enough. FP8 is data-free, so the quantization recipe above is the whole reproduce step — no calibration set needed.

Validation (GB10, transformers)

  • Text: coherent EN + ZH.
  • Image: accurately described a studio-podcast photo (cat + Shiba Inu, headphones, studio mics, latte-art mug).
  • Audio: understood a LibriSpeech clip (Mr. Quilter / apostle / middle classes).
  • Video: correctly described a night-street clip.

Credits

  • Base model: google/gemma-4-12B-it (Apache 2.0, Google DeepMind)
  • Quantization: llmcompressor + compressed-tensors
  • Quantized & benchmarked by coolthor on a DGX Spark (GB10)

Support

One-person effort on a single DGX Spark, no sponsor. If it saved you time, a coffee ☕ is appreciated.